SGLang v0.5.18 introduces default cache directory changes

The sgl-project/sglang v0.5.18 release, as detailed in the project's GitHub release notes, introduces numerous updates, including new models, optimizations, and breaking changes. Among the notable changes, all compiled-kernel caches are now consolidated under the SGLANG_CACHE_DIR environment variable. This affects Triton, FlashInfer, Inductor, DeepGEMM, and CUDA driver caches, requiring recompilation during the first launch after upgrading. Operators can opt to preserve the old cache locations by copying or symlinking directories or setting specific environment variables for individual caches.

Operators should ensure their systems have sufficient storage and write permissions for the new cache directory to avoid runtime errors during the initial recompilation. Additionally, pre-warming or volume-mounting the caches may mitigate performance impacts. This change reflects a broader trend in projects consolidating cache handling for improved manageability, but it may disrupt workflows relying on previous cache paths.